Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes

Adaptive Sampling to Reduce Epistemic Uncertainty Using Prediction Interval-Generation Neural Networks

AAAI 2025 · Probabilistic and scientific modeling › Uncertainty and Gaussian processes · approach: Epistemic versus aleatoric decomposition · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ASPINN, an adaptive sampling method that scores epistemic uncertainty from prediction-interval neural networks (DualAQD) and selects batch samples via a GP surrogate to reduce epistemic uncertainty across the input domain.

Problem

Reducing epistemic uncertainty across an entire input domain with few costly experiments, where existing uncertainty-driven sampling is mostly framed as active learning or optimization and cannot separate epistemic from aleatoric uncertainty cleanly.

Approach

A potential-epistemic-uncertainty metric built from neural prediction-interval bounds and the nearby observed points that fall inside them, combined with a GP surrogate whose covariance is scaled by this metric; an acquisition function picks the sampling location that maximizes the drop in total potential uncertainty over the domain.

Key results

• On the 1-D problems, ASPINN achieved the lowest AUUC: cos 97.26 vs 112.57 (MC-Dropout), 123.87 (GP), 113.39 (NF-Ensemble); hetero 85.95 vs 106.44 to 113.80; cosqr 17.13 vs 23.12 to 30.39.
• On the simulated field site, ASPINN AUUC was 496.85 vs 593.54 (GP), 614.68 (MC-Dropout), 730.80 (NF-Ensemble), with significant p-values for all comparisons.
• ASPINN's differences from all baselines were statistically significant at 0.05 in paired t-tests on the 1-D problems and the field site.

Measured effects reported in the paper

Surprising findings

• The standard GP baseline outperformed MC-Dropout and NF-Ensemble on the simulated field site's AUUC, while NF-Ensemble was the worst there.
• The authors argue that the common KL-divergence-based evaluation is inconsistent across methods, so they evaluate with an independent PI-generation network trained with DualAQD for all methods.

Limitations

• Does not handle multi-modal aleatoric noise; requires multiple PI handling for such cases.
• Computational cost grows with high-dimensional input spaces because all candidate locations must be evaluated.
• Evaluated only on synthetic 1-D problems and a simulated field site, not on real experimental data.

Experimental setup

Datasets: cos synthetic 1-D hetero synthetic 1-D cosqr synthetic 1-D simulated winter wheat field site

Benchmarks: cos hetero cosqr simulated field site

Metrics: PI_delta deviation from ideal 95% PI area under uncertainty curve (AUUC) paired t-test p-values

Baselines: MC-Dropout Gaussian process NF-Ensemble

Builds on: DualAQD Gaussian processes MC-Dropout normalizing flows ensembles

Closest papers by meaning

• General Uncertainty Estimation with Delta Variances (AAAI 2025 · similarity 0.87)
• Uncertainty Quantification with the Empirical Neural Tangent Kernel (NeurIPS 2025 · similarity 0.86)
• Epistemic Uncertainty Quantification For Pre-Trained Neural Networks (CVPR 2024 · similarity 0.85)
• Estimating Epistemic and Aleatoric Uncertainty with a Single Model (NeurIPS 2024 · similarity 0.85)
• A Study of Bayesian Neural Network Surrogates for Bayesian Optimization (ICLR 2024 · similarity 0.84)
• Quantifying Uncertainty in the Presence of Distribution Shifts (NeurIPS 2025 · similarity 0.84)